What problem does it solve? Setting up a lunch-ordering profile by intuition alone produces wrong rankings: intuitive answers like "most calories" or "flag my favorite protein" fail against real menu data, silently burying dishes the user would want or firing flags so often they become noise. ## Core Features & Use Cases - Menu-grounded interview: Fetches the live Lunsjkokkene menu first and asks concrete questions about real dishes, prices, proteins, allergies, categories, and portion sizes. - Profile generation: Writes a personal preferences.json file following the annotated preferences.example.json template, capturing hard vs. soft exclusions, category tiers, and repetition tolerance. - Reality-check validation: Runs the ranking against live menu data, shows top and bottom picks, and measures the Ukesmeny flag rate (targeting 20-35%) so miscalibrated weights are caught before saving. - Use Case: A user says their lunch suggestions are boring or wrong; the skill re-interviews them, adjusts the profile, and re-measures the flag rate until the weekly suggestions match their actual taste. ## Quick Start Ask the assistant to set up your lunch preferences by interviewing you about what you like to eat and then checking the suggestions against this week's actual menu.
